Method and Apparatus for Optimizing Tag of Point of Interest

1. A method for optimizing a tag of a point of interest, comprising: acquiring a set of points of interest and a set of tags of points of interest in the set of points of interest; generating a point of interest-tag matrix based on the set of points of interest and the set of tags of points of interest in the set of points of interest; extracting a feature of a point of interest-tag in the point of interest-tag matrix; inputting the feature of the point of interest-tag in the point of interest-tag matrix into a pre-trained ternary self-adaptive collaborative learning model, to obtain a point of interest-tag score matrix; and optimizing the set of tags of points of interest in the set of points of interest based on the point of interest-tag score matrix.

2. The method according to claim 1 , wherein the feature of the point of interest-tag comprises at least one of following items: an attribute feature of a point of interest, an image feature of the point of interest, an image feature of a tag, a tag feature of the point of interest, or a tag feature of the tag.

3. The method according to claim 2 , wherein the extracting a feature of a point of interest-tag in the point of interest-tag matrix comprises: determining, for a point of interest in the point of interest-tag matrix, historical access users or historical retrieval users of the point of interest; aggregating user images of the historical access users or the historical retrieval users of the point of interest to obtain an image feature of the point of interest; determining, for a tag in the point of interest-tag matrix, a point of interest belonging to the tag; and generating an image feature of the tag based on the image feature of the points of interest belonging to the tag; or determining, for a point of interest in the point of interest-tag matrix, an adjacent point of interest to the point of interest based on map retrieval data; statisticizing tag distribution of the adjacent point of interest to the point of interest to obtain a tag feature of the point of interest; determining, for a tag in the point of interest-tag matrix, a point of interest belonging to the tag; and generating the tag feature of the tag based on the tag feature of the point of interest belonging to the tag.

4. The method according to claim 3 , wherein the generating an image feature of the tag based on the image feature of the point of interest belonging to the tag comprises: computing an average value of the image feature of the point of interest belonging to the tag, for use as the image feature of the tag, and the generating the tag feature of the tag based on the tag feature of the point of interest belonging to the tag comprises: computing an average value of the tag feature of the point of interest belonging to the tag, for use as the tag feature of the tag.

5. The method according to claim 1 , wherein before the inputting the feature of the point of interest-tag in the point of interest-tag matrix into a pre-trained ternary self-adaptive collaborative learning model, to obtain a point of interest-tag score matrix, the method further comprises: integrating a pre-trained first machine learning model, a second machine learning model, and a prediction model to obtain the ternary self-adaptive collaborative learning model, wherein the first machine learning model is established based on points of interest in the point of interest-tag matrix, the second machine learning model is established based on tags in the point of interest-tag matrix, and the prediction model is established based on the points of interest in the point of interest-tag matrix.

6. The method according to claim 5 , wherein the first machine learning model, the second machine learning model, and the prediction model are obtained by following training: establishing a decomposition process from the point of interest-tag matrix to QR by non-negative matrix factorization, to obtain a decomposition loss function, wherein the QR is a result of orthogonal triangular decomposition of the point of interest-tag matrix, Q is an orthogonal matrix, and R is an upper triangular matrix; establishing a matching model from the point of interest to the tag, wherein the matching model is established based on a product of the first machine learning model and a transpose of the second machine learning model; optimizing the first machine learning model and the second machine learning model by cross entropy, to obtain a first loss function; obtaining a second loss function, a third loss function, and a fourth loss function based on the first machine learning model, the second machine learning model, and the QR; performing weighted summation on the first loss function, the second loss function, the third loss function, and the fourth loss function, to obtain a loss function of the first machine learning model and the second machine learning model; optimizing the prediction model by cross entropy to obtain a fifth loss function; obtaining a sixth loss function based on the prediction model and the QR; performing weighted summation on the fifth loss function and the sixth loss function, to obtain a loss function of the prediction model; and minimizing a weighted sum of the decomposition loss function, the loss function of the first machine learning model and the second machine learning model, and the loss function of the prediction model, and obtaining the first machine learning model, the second machine learning model, and the prediction model by training.

7. The method according to claim 6 , wherein the optimizing the first machine learning model and the second machine learning model by cross entropy, to obtain a first loss function comprises: computing the first loss function based on the point of interest-tag matrix, the first machine learning model, and the second machine learning model.

8. The method according to claim 6 , wherein the obtaining a second loss function, a third loss function, and a fourth loss function based on the first machine learning model, the second machine learning model, and the QR comprises: computing a column sum norm of a difference between a product of the first machine learning model and the second machine learning model, and the QR, to obtain the second loss function; computing a spectral norm of a difference between a product of the first machine learning model and a transpose of the first machine learning model, and a product of the QR and a transpose of the QR, to obtain the third loss function; and computing a spectral norm of a difference between a product of the second machine learning model and a transpose of the second machine learning model, and a product of the transpose of the QR and the QR, to obtain the fourth loss function.

9. The method according to claim 6 , wherein the optimizing the prediction model by cross entropy to obtain a fifth loss function comprises: computing the fifth loss function based on the point of interest-tag matrix and the prediction model.

10. The method according to claim 6 , wherein the obtaining a sixth loss function based on the prediction model and the QR comprises: computing a spectral norm of a difference between the prediction model and the QR, to obtain the sixth loss function.

11. The method according to claim 1 , wherein the optimizing the set of tags of points of interest in the set of points of interest based on the point of interest-tag score matrix comprises: adding, for a point of interest-tag score in the point of interest-tag score matrix, in response to the point of interest-tag score being higher than a first preset score, and a set of tags of points of interest corresponding to the point of interest-tag score excluding a tag corresponding to the point of interest-tag, the tag corresponding to the point of interest-tag into the set of tags of points of interest corresponding to the point of interest-tag score; and deleting, in response to the point of interest-tag score being lower than a second preset score, and the set of tags of points of interest corresponding to the point of interest-tag score including the tag corresponding to the point of interest-tag, the tag corresponding to the point of interest-tag from the set of tags of points of interest corresponding to the point of interest-tag score.

12. An apparatus for optimizing a tag of a point of interest, comprising: at least one processor; and a memory storing instructions, wherein the instructions when executed by the at least one processor, cause the at least one processor to perform operations, the operations comprising: acquiring a set of points of interest and a set of tags of points of interest in the set of points of interest; generating a point of interest-tag matrix based on the set of points of interest and the set of tags of points of interest in the set of points of interest; extracting a feature of a point of interest-tag in the point of interest-tag matrix; inputting the feature of the point of interest-tag in the point of interest-tag matrix into a pre-trained ternary self-adaptive collaborative learning model, to obtain a point of interest-tag score matrix; and optimizing the set of tags of points of interest in the set of points of interest based on the point of interest-tag score matrix.

13. A non-transitory computer readable medium, storing a computer program thereon, wherein the computer program, when executed by a processor, implements the method according to claim 1 .